Abstract

Foundation models are transforming structured data learning much like large language models did for text. In this talk, I will present our new foundation models, Mitra and Chronos-2, which demonstrate how synthetic pretraining and in-context learning (ICL) enable models to generalize across diverse tabular and time-series tasks without task-specific training. Mitra curates a principled mixture of synthetic priors to achieve state-of-the-art performance on tabular classification and regression, while Chronos-2 introduces group attention to unify univariate, multivariate, and covariate-informed forecasting. Together, they illustrate a new paradigm where the design of synthetic data priors and ICL mechanisms, rather than per-task fine-tuning, drives generalization and scalability across structured domains.

Bio: Xiyuan Zhang is an Applied Scientist at Amazon Web Services working on machine learning for structured data (time series, tabular), especially on pre-training and multimodal analysis. She is the lead author of Mitra, the most downloaded tabular foundation model on HuggingFace, and co-author of Chronos, the most downloaded time series foundation model on HuggingFace. Xiyuan earned her PhD in Computer Science from the University of California, San Diego. She is a recipient of the Qualcomm Innovation Fellowship and has been recognized as a Cyber-Physical-System (CPS) Rising Star.